(when (seq coll)
...)
(Note: this page describes the changes made in the last large update to sequences, when Rich was also exploring an alternative called streams. Treat this page as a useful historical record of these design decisions, rather than as reference documentation.)
In working on streams a few things became evident:
stream code is ugly and imperative
even when made safe, still ugly, stateful
streams support full laziness
this ends up being extremely nice
Integrating streams transparently (i.e. not having both map and map-stream) would require a change, to relax the contract of the core sequence functions (map, filter etc)
If I am going to do that, could I achieve the same full laziness while keeping the beautiful recursive style of Clojure?
while being substantially compatible with existing code
Yes!
but - ideally some names should change
Originally modeled on Common Lisp’s cons
Either you have a seq, with a valid first, or nothing (nil)
A seq is like a logical cursor
rest is fundamentally eager
returns another seq or nil
needs to determine if there is any more in order to determine if return value is nil
lazy-cons can delay the calculation of first/rest, but not the determination if there is a rest
determining that often requires 'pulling' on an inner seq, reducing effective laziness
sequence functions currently return a seq or nil
The eagerness of rest means the sequence functions are not fully lazy, they need to at least determine if there is a first.
Changed: (rest aseq) - returns a possibly empty seq, never nil
calls seq on arg if not already a seq
the returned seq may be empty
will print (), but not a single sentinel object
never returns nil
currently not enforced on 3rd-party seqs
a (possibly) delayed path to the remaining items, if any
Changed: seqs can be empty
always an ISeq
Changed: the seq function - no longer an identity for ISeqs
still returns either a seq or nil
(seq aseq) → no longer an identity, if aseq empty returns nil
still works on nil
the first function doesn’t change
calls seq on arg if not already a seq
returns first item
still works on nil
New: the next function does what rest used to do
returns the next seq, if any, else nil
calls seq on arg if not already a seq
(next aseq) === (seq (rest aseq))
works on nil
Changed: seq?
(seq? ()) → true
Changed: Sequence fns (map, filter etc) return seqs, but not nil
You’ll need to call seq on their return value in order to get a seq/nil
seq also serves as test for end, already idiomatic
(when (seq coll)
...)allows full laziness
doesn’t support nil punning
since sequence fns no longer return seq/nil
Goodbye lazy-cons, hello lazy-seq
lazy-cons is gone
new laziness macro - lazy-seq
takes a body that yields a seq, nil or anything seq-able
returns a logical collection that implements seq by calling the body
invokes the body only the first time seq is called on it, caches result
will call seq on the body’s return value if not already a seq or nil
The net effect is the creation of a virtual collection that does no work until seq is called upon it - fully delayed
Supports all collection ops
Can be empty - e.g. calling seq on it can return nil
when empty will print as ()
lazy-seq goes at top level of lazy sequence function
instead of nested lazy-cons
inside, use a normal cons call
won’t be created until needed
if consuming another seq, use rest instead of next
The old way:
(defn map
([f coll]
(when (seq coll)
(lazy-cons (f (first coll)) (map f (rest coll)))))
...The new way:
(defn map
([f coll]
(lazy-seq
(when-let [s (seq coll)]
(cons (f (first s)) (map f (rest s))))))
...Note the use of when-let, which grabs the seq once, for subsequent use in first and rest, even though first/rest call seq on their argument. This has a performance benefit in this new model.
One of the nice things about CL’s cons using nil for end-of-list is that, when coupled with nil’s testability in conditionals, cons-returning functions could be used like predicates. Now only seq and next can be used in that manner - map, filter etc cannot. Note that much of the economy of the seq/nil dyad still applies, e.g. the use of when in map above.
If you are extending ISeq you’ll need to support ISeq.more() (the underpinnings of rest). Fortunately, most ISeq extenders derive from ASeq, which defines more() in terms of next. If you derive your seq from ASeq, don’t define more(), use the version supplied by ASeq. Just rename your rest() method to next().
To move to the new model you’ll need to take the following steps, in this order:
Rename all your calls to rest to call next
If you were defining your own lazy sequence functions, using lazy-cons, switch them over to lazy-seq using the recipe above. Make sure to call rest and not next in your recursive call.
Audit your code for nil-punning. The lazy branch has supports compilation in a debug mode that asserts if you try to test the truth value of a lazy sequence in a conditional, and will throw an exception if you do. Just build clojure like so:
ant -Dclojure.assert-if-lazy-seq=true
Then, nil puns like the following will throw exceptions:
(when (filter neg? [1 2]) :all-pos)
(not (concat))
(if (rest (seq [])) 1 2)
In all cases you can fix a nil pun by wrapping the sequence with a seq call:
(when (seq (filter neg? [1 2])) :all-pos)
-> nilAfter you are done, rebuild without the flag, as it will slow things down.
Recursively defined lazy sequence functions are elegant and easy to understand. They can be very memory efficient, allowing you to work with data sources that might not fit in memory, because only the part of the data structure in current use need be in memory. It could be tricky at times to determine which parts were currently in use, as they might still be referenced by local variables. Clojure does local-variable clearing on tail calls to ensure that no lingering references remain on the stack, but there was one remaining case - closed-over locals, that was difficult to control, especially when using a macro like lazy-seq which creates a closure on your behalf.
Consider the original, not fully lazy, definition of filter:
(defn filter
"Returns a lazy seq of the items in coll for which
(pred item) returns true. pred must be free of side-effects."
[pred coll]
(when (seq coll)
(if (pred (first coll))
(lazy-cons (first coll) (filter pred (rest coll)))
(recur pred (rest coll)))))By recurring to the fn itself, it is effectively erasing the coll argument each iteration, so it looks like it wouldn’t retain coll while skipping elements not matching the predicate. The problem is that sometimes the call to filter is in the lazy-cons, which expands into a closure that closes over coll, thus retaining it while the looping occurs, and there is nothing the called function can do about it. This means that expressions like:
(filter #(= % 20) (map inc (range 10000000)))could cause out of memory exceptions. The only way to avoid it was to rewrite filter using mutation. Bleh.
The new filter looks like this:
(defn filter
"Returns a lazy sequence of the items in coll for which
(pred item) returns true. pred must be free of side-effects."
[pred coll]
(let [step (fn [p c]
(when-let [s (seq c)]
(if (p (first s))
(cons (first s) (filter p (rest s)))
(recur p (rest s)))))]
(lazy-seq (step pred coll))))The body of the old filter has been put in a helper fn, and lazy-cons replaced with cons, then the whole call is wrapped in a lazy-seq, following the recipe above. However lazy-seq also creates a closure which closes over coll. Without some enhancement, this filter, while lazier, will have the same memory footprint as the old. The new lazy branch contains a compiler enhancement for this and similar scenarios. lazy-seq and delay both perform closed-over local clearing on the tail call of their body, ensuring no references remain in the closure itself when the tail-call executes. They can do this because they cache the results, and thus know the closure will be invoked only once. Thus the lazy branch has no problems with the filter expression above, and you can use similar techniques to control memory usage in your own lazy functions.